Image Feature Exploration

I manually picked 12 representatives from each category:

chicken_0116 chicken_0131 (so hard to distinguish) chicken_0181 chicken_0218 chicken_0264 chicken_0306 chicken_0383 chicken_0761 chicken_0646 chicken_0599 (only a leg) chicken_0891 (dark color) chicken_0999

dog_0010 dog_0021 (typical hair) dog_0023 dog_0033 dog_0095 dog_0146 (only a leg) dog_0147 dog_0342 (smooth surface) dog_0404 dog_0432 dog_0608 dog_0887

Some outliers in the whole dataset exist and may result in prediction errors.

Some ideas: - Extract surface texture which stands out with curvy grains (? hard to realize) - Color RGB feature (success) - Color HSV feature

Here I pick 5 images (3 fried chickens and 2 dogs) to do a image exloration demo.

Change to Greyscale, blue and red channels

file.bw=list()
file.bl=list()
file.red=list()
for (i in 1:5){
  file.bw[[i]]=channel(filename[[i]],mode="grey")
  display(file.bw[[i]])
}

for (i in 1:5) {
  file.bl[[i]]=channel(filename[[i]], mode='asblue')
  display(channel(filename[[i]], mode='asblue'))
}

for (i in 1:5) {
  file.red[[i]]=channel(filename[[i]], mode='asred')
  display(channel(filename[[i]], mode='asred'))
}

Extract the texture by high contrast on blue channel, which really shows difference to me between fried chicken and dogs.

file.contrast=list()
for (i in 1:5){
  file.contrast=file.bl[[i]]*5
  display(file.bl[[i]]*5)
}

Filtering: high pass

filter_high <- matrix(1, nc = 3, nr = 3)
filter_high[2, 2] <- -8
file.hi=list()
for (i in 1:5){
  file.hi <- filter2(filename[[i]], filter_high)
  display(file.hi)# high pass
}

Adaptive thresholding

for (i in 1:5){
  file.bw[[i]]=channel(filename[[i]],mode="grey")
  img_seg1 <- thresh(file.bw[[i]], w=60, h=60, offset=0.06)
  display(img_seg1, all=TRUE)
}

Outline Analysis

#Oriented Contour ### didn't see any use
img_leaf=list()
img_leaf1=list()
oc=list()
for (i in 1:5){
  img_leaf[[i]] <- resize(filename[[i]], 128, 128)
  img_leaf[[i]] <- channel(img_leaf[[i]], mode="gray")
  img_leaf1[[i]] <- thresh(img_leaf[[i]], w=50, h=50, offset=0.05)
  display(img_leaf1[[i]])}

for (i in 1:5){
  oc[[i]] <- ocontour(bwlabel(img_leaf1[[i]]))
  plot(oc[[i]][[1]], type='l');points(oc[[i]][[1]], col=2)
}

#Local Curvature
lc=list()
i=list()
neg=list()
pos=list()
for (k in 1:5){
  lc[[k]] <- localCurvature(x=oc[[k]][[1]], h=11)
  i[[k]] <- lc[[k]]$curvature >= 0
  neg[[k]] <- array(0, dim(img_leaf1[[k]]))
  pos[[k]] <- neg[[k]]
  pos[[k]][lc[[k]]$contour[i[[k]],]+1]  <- lc[[k]]$curvature[i[[k]]]
  neg[[k]][lc[[k]]$contour[!i[[k]],]+1] <- -lc[[k]]$curvature[!i[[k]]]
  display(10*(rgbImage(pos[[k]], , neg[[k]])), title = "Image curvature")
}

Color Features from RGB

By discretizing the RGB values, we can get a set of color features. We subdivide the pixel values in each color channel into multiple bands of equal width. Then the counts of pixels in correponding bins consist of a set of color features that characterizes the color distribution of the image.

str(rgb_feature)

List of 5
 $ : num [1:800] 0.01379 0.02122 0.02635 0.02134 0.00368 ...
 $ : num [1:800] 0.00301 0.01485 0.01819 0.02569 0.02503 ...
 $ : num [1:800] 7.88e-06 1.38e-03 8.54e-03 9.58e-03 3.15e-04 ...
 $ : num [1:800] 0.001236 0.005528 0.006125 0.000306 0 ...
 $ : num [1:800] 0.01459 0.01061 0.01105 0.00498 0.00152 ...